A Real Claude Cowork Walkthrough: 4,000 Accounts to Pipeline

Abstract advice about agentic sales is easy to nod along to and hard to act on. So this post is a single, concrete walkthrough: a rep inherits a 4,000-account book on a Monday and, by using Claude Cowork deliberately, has prioritized it, researched the top tier, and shipped reviewed outreach by the end of the week. Every step below is something you can actually reproduce, including where the agent gets things wrong and what the human does about it.

The point is not that Cowork magically closes deals. The point is the operating loop — specify, generate, verify, ship — applied to a book no human could process by hand, and the specific decisions that make it work instead of producing 4,000 mediocre emails.

The starting state: a messy 4,000-account list

The book arrives as a CRM export: company names, some firmographics, partial contact data, a scatter of old notes. Roughly a third of the records are stale. No human is going to read all of it. The temptation is to start at the top of the alphabet and grind, which guarantees the best accounts get reached in month three. The first job is therefore not outreach — it is triage, and triage is exactly the kind of consistent, high-volume judgment Cowork is good at when given a clear specification.

Before touching the agent, the rep writes down what a tier-1 account looks like for this product this quarter: company size band, a recent triggering signal, evidence of the pain the product solves, and a clean contact. This written definition is the specification the whole workflow hangs on. Vague definitions produce vague tiering; a sharp one produces a list the rep can trust.

Step one: connect tools and prioritize

Cowork connects to the CRM and an enrichment data source through MCP connectors — read-only, so the agent can pull everything but cannot yet change anything. The rep gives it the tier-1 specification and asks for the 4,000 accounts sorted into tiers with a one-line reason per account. This is a read-and-draft task: maximum usefulness, near-zero blast radius, because nothing is being written or sent.

flowchart TD
  A["Rep writes tier-1 spec"] --> B["Cowork pulls 4,000 records via MCP"]
  B --> C["Agent tiers & gives reasons"]
  C --> D["Rep sample-audits tiering"]
  D -->|Spec wrong| A
  D -->|Good| E["Research skill briefs tier-1"]
  E --> F["Draft skill writes openers"]
  F --> G["Rep edits voice & sends"]
  G --> H["Replies logged back to CRM"]

The agent returns the tiering in minutes. The rep does not accept it blindly — they sample-audit, reading twenty accounts across the tiers. Two are mis-tiered because the specification under-weighted a signal. That feedback goes back into the spec, the agent re-runs, and now the tiering holds up. This loop — generate, sample, fix the specification, regenerate — is the heartbeat of the whole walkthrough and it costs minutes, not days.

Step two: research the top tier at volume

With perhaps 300 tier-1 accounts identified, the rep invokes a reusable research skill — a Cowork skill that encodes "how we research an account": what to look for, which sources to weigh, and the exact shape of the output brief. Because it is a skill rather than an ad-hoc prompt, every brief comes back in the same structure, which makes them fast to read and easy to compare.

Here is where a definition earns its place: an Agent Skill is a reusable bundle of instructions and resources that an agent loads when relevant, so a workflow like account research runs the same way every time instead of depending on how each request happens to be phrased. The research skill turns "research these 300 accounts" from an impossible afternoon into a batch the agent works through while the rep does something else, returning structured briefs the rep then skims.

The rep reads the briefs critically. Several misread a funding event or attached an outdated org chart. The rep corrects those in place and notes the pattern — the skill is over-trusting one stale source — so RevOps can tighten the skill later. The errors are caught because the briefs are only read by a human, never acted on automatically. Low blast radius, by design.

Step three: draft, verify, and ship

Now a separate drafting skill turns each researched brief into a first-touch opener that references the real signal the research surfaced. The agent produces 300 tailored drafts. They are good — specific, on-topic — but they are drafts, not sends. The rep moves through them as an editor: fixing voice, killing the three that misfire, personalizing the handful going to marquee accounts. A draft that took fifteen minutes to write by hand now takes ninety seconds to verify and polish.

Sending is the one high-blast-radius step, so it is gated. The rep sends in reviewed batches rather than firing all 300 at once, and the first batch's replies are watched closely as a final quality check. Replies and outcomes log back to the CRM — through a narrowly-scoped write connector, separate from the read connectors used earlier — closing the loop so the next prioritization pass learns from what actually landed.

What shipped, and what made it work

By Friday the rep has a trustworthy tiering across all 4,000 accounts, structured research on the top 300, and a few hundred reviewed, on-brand openers sent in controlled batches. No human could have produced that by hand in a week. But the result did not come from "the AI doing sales." It came from three deliberate choices: a sharp written specification, reusable skills so quality compounds, and a human editing at the points of judgment while the agent did the volume.

The failures along the way — mis-tiering, stale research sources, a few off drafts — were all caught cheaply because every error happened in a draft a human read, never in an unattended action. That is the whole trick. Put the agent where it has leverage and the human where they have judgment, and a 4,000-account book becomes a Monday-to-Friday project instead of a quarter-long grind.

Frequently asked questions

How long does the first full pass realistically take?

For a rep who already has the specifications and skills in place, a first prioritize-research-draft pass over a large book is a few days of focused work, most of it verification rather than origination. The slow part is the very first time, when you are still writing the specs and skills; after that the same machinery runs far faster on the next book.

What goes wrong most often in this workflow?

Over-trusting the agent's output. The walkthrough works because every step that could cause harm — tiering, research, sends — is sampled or gated by a human. Teams that skip the sample-audit and ship agent output directly get volume, but they also ship the agent's mistakes at volume.

Why use skills instead of just prompting each time?

Consistency and reuse. A skill makes research or drafting run the same way every time and lets a good approach, written once, serve the entire book and every future rep. Ad-hoc prompts give you results that vary with phrasing and never accumulate into a shared asset.

The same loop, on your phone lines

CallSphere runs this specify-generate-verify loop for voice and chat: agentic assistants that answer every call and message, pull and write data mid-conversation, and book work 24/7. See the live version at callsphere.ai.

---

Source & attribution: This is an independent, original explainer inspired by Anthropic's coverage on the Claude blog. Claude, Claude Code, Claude Cowork, Claude Opus, and the Model Context Protocol are products and trademarks of Anthropic. CallSphere is not affiliated with or endorsed by Anthropic.